From Drug Action to a Patient-Relevant Outcome
A drug can bind its target and change a biological measurement without improving how patients feel or function. The key question is not merely whether a biomarker responds to treatment, or whether it correlates with disease severity. It is whether the biomarker helps explain a plausible causal path from drug action to a clinical outcome.
The earlier material on evidence chains and counterfactual questions provides useful framing: a mechanistic explanation proposes what would change if the drug were given, and what would happen to the clinical outcome if a link in that chain were altered. A biomarker can make that chain observable, but measurement alone does not establish the link.
A proposed bridge might be written as drug exposure → target engagement → change in a disease-relevant process measured by a biomarker → change in symptoms. Each arrow is a distinct claim. For example, target engagement may be measured reliably while the target has little influence on the disease process; a downstream biomarker may change while a compensating pathway preserves symptoms; or the biomarker may reflect disease severity without causing it.
What Makes a Biomarker a Plausible Intermediate?
A useful starting point is to specify the proposed causal order and what each measurement represents. The drug must affect the biomarker at a relevant exposure and time. The biomarker must represent a process with a credible biological connection to the clinical outcome. Its change should precede the outcome change when the mechanism implies that sequence. Finally, evidence should show that treatment effects on the biomarker and outcome are consistent with the proposed link—not just that the two measurements vary together.
In a simplified causal diagram, treatment \(A\) affects a biomarker \(M\), which in turn affects an outcome \(Y\): \(A \rightarrow M \rightarrow Y\). There may also be a direct path from \(A\) to \(Y\), and other factors may influence both \(M\) and \(Y\). Randomizing \(A\) helps estimate the effect of treatment, but it does not automatically randomize \(M\). Thus, an association between \(M\) and \(Y\) can still reflect disease severity, another biological process, or a treatment-related factor rather than a causal effect of the biomarker-linked process.
A simple model helps make the bridge claim precise. Suppose the biomarker and outcome are measured so that larger values have a defined direction, and consider:
Here \(a\) describes the treatment effect on the biomarker, \(b\) describes the biomarker-outcome relation after accounting for treatment, and \(c'\) is the remaining treatment effect on the outcome in this model. Substituting the first equation into the second gives:
So, under this additive linear model, the total treatment effect \(c\) is \(c' + ab\). The product \(ab\) is the model’s mediated component. This identity is an algebraic decomposition; by itself, it does not prove that \(M\) causes \(Y\). A causal interpretation requires, among other conditions, a well-defined treatment, correct time ordering and model form, and no unmeasured confounding of the biomarker-outcome relation. It also requires care when treatment changes a factor that then confounds that relation. Without these conditions, \(ab\) is a model-based association, not an established mediated effect.
Worked Example: A Plausible Inflammatory Intermediate
Hypothetical data. In a randomized study, a drug lowers an inflammation index by 4 units compared with control. The average symptom score falls by 8 points, with lower scores indicating fewer symptoms. In a separate mechanistic analysis, a 1-unit decrease in the index is associated with a 1.5-point decrease in symptom score after accounting for treatment.
Represent a decrease in the index as a change of \(-4\) units and a decrease in symptom score as \(-8\) points. The model-implied mediated component is:
The observed total treatment effect is \(-8\) points. Under the stated linear decomposition, the remaining component is:
The model therefore attributes 6 of the 8 points of average improvement to the biomarker-linked component, leaving 2 points not accounted for by that component. The arithmetic is consistent with partial mediation, not proof of it. The result is more credible if the index changes before symptoms, the biology supports the proposed link, and evidence from interventions that alter the process also supports a corresponding clinical effect. If the biomarker-outcome relation is confounded, the calculated \(-6\) points may not be a causal contribution.
Association Can Arise Without a Biomarker-to-Outcome Effect
A biomarker and clinical outcome can move together because both reflect a third factor. Disease severity is a common example: it can raise a biomarker and worsen symptoms, even if intervening on the biomarker itself would not improve symptoms. This is why overall correlation is not enough to support a mechanistic bridge.
Worked Example: Severity Creates a Misleading Association
Hypothetical observations. Consider 16 patients split into two equally sized groups. In the eight patients with severe disease, the mean biomarker level is 10 units and the mean symptom score is 8 points. In the eight with moderate disease, the corresponding means are 2 units and 2 points. Within each severity group, suppose the biomarker and symptoms do not vary together.
Across the two groups, the difference in mean biomarker level is \(10-2=8\) units, and the difference in mean symptom score is \(8-2=6\) points. The combined data therefore show that higher biomarker levels accompany worse symptoms. But the pattern is fully explained by the severity-group contrast: severe disease is associated with both measurements. Within either group, the stated data provide no evidence that patients with higher biomarker values have worse symptoms.
The group-level association does not establish that reducing the biomarker would reduce symptoms. To support that conclusion, one would need evidence that separates the biomarker-linked process from severity and other common causes, and that tests the proposed causal path. A biomarker may still be useful for describing disease state or predicting outcomes; those uses are different from establishing it as a mechanistic intermediate.
Temporal order is a useful diagnostic, but not a complete solution. A biomarker that changes only after symptoms improve is unlikely to explain the initial improvement. Conversely, a biomarker that changes first may still be a bystander if both changes are driven by another treatment effect. Biological plausibility and timing strengthen a hypothesis; neither alone establishes causality.
Target Engagement Is Not the Same as Clinical Bridging
Worked Example: Strong Target Engagement, Little Clinical Change
Hypothetical data. A drug produces 80% target occupancy at a low dose and 95% at a high dose, compared with 0% on placebo. The average clinical improvement is 1.0 symptom point at the low dose and 1.2 points at the high dose; the prespecified clinically important improvement is 4 points.
Both active doses show substantial target engagement, but neither reaches the 4-point clinical benchmark: \(1.0<4\) and \(1.2<4\). Increasing occupancy by 15 percentage points is accompanied by only \(1.2-1.0=0.2\) additional symptom points of improvement. This hypothetical pattern supports target engagement, but does not establish that the measured occupancy bridges to clinically meaningful benefit.
Possible explanations include a weak connection between the target and the disease process, saturation of a downstream response, or compensation through another pathway. The data do not distinguish these explanations on their own. They do, however, warn against treating occupancy as a substitute for the clinical outcome.
The right interpretation depends on the intended use. A biomarker can be useful for confirming target engagement, characterizing a biological response, or helping predict which patients may respond, while still being unsuitable as a replacement for clinical outcomes. A surrogate claim is especially demanding because it must predict treatment effects on outcomes, not merely individual patients’ outcome levels.
Practical Assessment and Common Failure Modes
When assessing a proposed biomarker bridge, state the claim at the level the evidence supports. Is the marker a measure of target engagement, a response marker, a prognostic marker, a candidate intermediate, or a validated surrogate in a defined context? These labels make different claims and should not be used interchangeably.
- Check the proposed sequence. Confirm that drug action, biomarker change, and clinical response are measured on a timeline compatible with the mechanism.
- Check what the assay measures. A measured signal may be an indirect proxy for a biological process, and assay variability or sampling location can weaken its interpretation.
- Look for competing explanations. Severity, other pathways, or treatment-related changes may affect both biomarker and outcome.
- Compare treatment effects, not only correlations. A biomarker that tracks outcome in one group may fail to predict how different treatments affect outcome.
- Limit the claim to its context. Evidence may apply only to a particular disease, population, intervention, biomarker definition, and time horizon.
A common mistake is to say, “The biomarker is correlated with symptoms and the drug changes the biomarker, so changing the biomarker must improve symptoms.” The conclusion does not follow: the correlation may be confounded, and the treatment may affect symptoms through another route. The correction is to treat the biomarker as a candidate link, state the causal assumptions, and seek evidence that tests the link itself.
A mechanistic model can make the proposed bridge explicit and expose which measurements or assumptions carry the conclusion. It cannot turn an unsupported link into evidence merely by including it as a model component. If the biomarker does not reliably represent the relevant process, or the process is not causally connected to the outcome in the context of interest, additional model detail will not resolve that gap.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A drug changes a target-occupancy marker, but no clinical outcome improves. What does this support, and what does it not establish?
- In the linear model above, treatment changes a biomarker by \(-3\) units and the biomarker-outcome coefficient is 2 outcome points per unit. What is the model-implied mediated component?
- Why can a biomarker-outcome correlation across patients be misleading if disease severity affects both?
- A biomarker changes after clinical improvement begins. What concern does this raise about interpreting it as the cause of the initial improvement?
- What additional claim is made by calling a biomarker a validated surrogate rather than a candidate mechanistic intermediate?
Solutions. (1) It supports target engagement or a biological response, depending on what the marker measures; it does not establish a clinical bridge or benefit. (2) \(ab=(-3)(2)=-6\) outcome points, a model-implied component whose causal interpretation still depends on the model assumptions. (3) Severity can be a common cause of both measurements, creating an association without a causal effect of the biomarker-linked process on outcome. (4) The timing is inconsistent with that biomarker change causing the onset of improvement, though it could relate to later response. (5) It claims that treatment effects on the biomarker reliably predict treatment effects on the clinical outcome in a specified context, not merely that a plausible mechanism or association exists.